There's a core problem this analysis overlooks: OpenAI and Anthropic don't have a moat. The Chinese labs are consistently able to replicate their LLM capabilities a few months after the fact, and then release open-weight models a few months later...

The only way for "Big AI" to become a thing is for them to establish a moat, and right now the only path to that appears to be achieving regulatory capture in the US, which is a fickle and unstable state of affairs.

You are not allowed to use any model from Chinese labs at American companies. Even from other companies like Mistral it is hard to get access, since enterprises require their own instances on vertex/bedrock/azure, and to run alternatives you have to go through multiple layers of management. In practice, no employee will proactively advocate for alternative models and the oligopoly is written in the wall, specially since the investments are circular anyway.

> "release open-weight models a few months later"

They are going to want to profit from their resources expended eventually and when they do, that will end.

That's why they're so keen to build out all these data centers; the massive capex is their only conceivable moat. Whether all those additional parameters actually make a competitive difference is still an open question.

Big AI is a thing. Insane money and construction projects makes the government invested in it one way or the other. It can't fail or it'll cause economic shock. It's a play to normalize it and make everyone depend on it before people question how they trained it all on their data and now drive them out of their jobs.

If they ask for a moat they will get it, the government can simply require licensing for hardware required to run any model comparable to the cloud ones (for "national security"). good local models will only be used by .01% who are crazy neckbeards running servers in their basement and even then they will suck compared to cloud ones.

We're in the Pets.com era of AI for sure. There is 0 guarantee the current providers will be around in 5-10 years. I keep thinking they will have to lean in to government protection to do so.

If your government weaponize AI against you it doesn't matter if the AI is American, Chinese or Uzbek

Memory might be a moat. A cynical view is this is why openAI is buying up all the capacity for HBM. If they have all the memory then we plebs won't be able to afford hardware to run local models. So we instead we have to rent their models basically just to get access to the memory.